Page 2 of 2
Re: Not Getting desired points from Project 14180 WUs
Posted: Tue Sep 03, 2019 4:12 pm
by bruce
Unfortunately there's really nothing that can be done about it from the server's perspective so we have to guess what it might be ... until we figure out what halps. Since it seems to be an occasional error (and as Joe_H has said, it's not a problem with the WU itself) this is going to be tricky. My best guess is that you might be able to minimize/remove the problem by underclocking memory on the GPU identified as running in software slot 3.
Re: Not Getting desired points from Project 14180 WUs
Posted: Wed Sep 04, 2019 3:06 am
by Luscious
Tricky yes because even if one card is bad (highly unlikely), it doesn't explain why system wide my PPD is performing below where it should and turning in fewer WU, and why WU coming from the other work server run absolutely fine.
But I am curious as well as to why the ETA in these projects is being misreported and taking 14+ hours in some cases such as 14191 (18,1,38) - it is rare for me to get a WU that takes more than 6-8 hours to complete (most finish in under 5). There is obviously a link between ETA, Estimated Credit and TPF in FAHControl, or something else is broken where the wrong Estimated Credit being reported is causing all the WU to run in limp mode.
I will continue to look for errors in my log and if I still don't see any change do a reboot in the coming days.
Re: Not Getting desired points from Project 14180 WUs
Posted: Wed Sep 04, 2019 4:18 pm
by bruce
The TPF is measured (for a small group of frames). They are averaged and projected to give an estimated completion time. The estimated points are based on the actual time required.
All those calculation are done locally so if your GPU is running slow, they'll all report that fact.
The actual points are measured by the server's clock which probably matches the local clock in your computer.
The actual speed that your GPU completes the calculations depends on a number of factors in your computer, including things like the number of shaders and their speed (effectively equivalence to GFLOPS.) Most of the actual GFLOPS consumed are SP calculations, but a small percentage are DP calculations. A small percentage of the time is spent moving data across the PCId bus and the internet. Another small percentage depends on the CPU speed and main RAM.